在win10平台上，异步io操作何时会在asio中调用on_completion？

Question

在asio_1.12.1 win_iocp_socket_service_base.ipp中，我在这种模式下发现了许多类似的代码：

start_async_action(...){//for example: start_send_op、 start_connect_op
    ...
    iocp_service_work_started();
    ...
    int result = ::WSA_IO_OPERATION(...); //such as WSASend、WSARecv
    DWORD last_error = ::WSAGetLastError();
    if (result != 0 /*or !result */ && last_error != WSA_IO_PENDING)
        iocp_service_.on_completion(...);  //means IO completed immediately
    else
        iocp_service_.on_pending(op); 

注意：它是iocp_service_.on_completion(...)，与op->on_complete(...)中的do_one

在多次跟踪程序后，始终调用on_pending，但从未看到调用on_completion。如何使场景调用此功能？ 我认为当IO操作立即完成时，它将调用on_completion，这可能会导致GetQueueCompletionStatus被绕开，因此on_completion进入了回调，消息是由PostQueueCompletionStatus发布的。

Ps：我在libuv中找到了代码，通过WSA操作标志绕过GetQueueComletionStatus，但在asio中不习惯调用on_completion：

result = WSARecv(handle->socket,
               (WSABUF*)&buf,
               1,
               &bytes,
               &flags,                /*flag for with or without IOCP */
               &req->u.io.overlapped,
               NULL);

if (UV_SUCCEEDED_WITHOUT_IOCP(result == 0)) {
    /* Process the req without IOCP. */
    handle->flags |= UV_HANDLE_READ_PENDING;
    req->u.io.overlapped.InternalHigh = bytes;
    handle->reqs_pending++;
    uv_insert_pending_req(loop, (uv_req_t*)req); }